Source-linked AI summary
WIP: Energy-Efficient LLM-Based Serving Cluster Formulation in Cell-Free Massive MIMO
Marcin Hoffmann, Paweł Kryszkiewicz
TL;DR
Selecting which base stations should jointly serve each user is central to improving energy efficiency in user-centric cell-free massive MIMO networks. This paper uses an LLM-based AI agent for dynamic serving-cluster formulation and reports up to 32% higher energy efficiency than the baseline in realistic simulations.
Problem
Serving-cluster formulation must determine which base stations jointly serve each user to improve energy efficiency in user-centric cell-free massive MIMO networks.
Method
An LLM-based AI agent performs dynamic serving-cluster formulation using reinforcement-learning interaction and network energy efficiency as its reward.
Results
Up to 32% higher energy efficiency was achieved than the baseline, with no decrease in quality of service for users under the best or worst radio conditions.
Takeaways & Limitations
LLM-based dynamic serving-cluster formulation can improve energy efficiency in realistic UCCF MMIMO simulations while preserving reported user quality of service.
Abstract
from arXiv · showhide
One way to increase the Energy Efficiency (EE) of 6G wireless networks is to utilize existing network infrastructure more efficiently. This can be achieved by introducing User-Centric Cell-Free Massive Multiple-Input-Multiple-Output (UCCF MMMIMO), which allows for simultaneously serving a single user by multiple Base Stations (BSs). From this perspective, the key challenge is to decide which BSs should serve a given user, known as the Serving Cluster Formulation (SCF). In this paper, we propose to deal with this problem by using an Artificial Intelligence (AI) agent based on a Large Language Model (LLM), targeting improvement of EE. We evaluated the proposed AI agent using a complex, 3D Ray Tracer-based, cellular network simulator, comparing a few GPT models and state-of-the-art algorithms. The results show up to 32% gain in EE of the proposed AI agent compared to the baseline.
I. Introduction
UCCF MMIMO is presented as a technology for improving 6G energy efficiency by coordinating multiple BSs, while the paper proposes an LLM-based AI agent for serving cluster formulation. The agent is evaluated in a realistic 3D Ray Tracer-based UCCF-MMIMO simulator to assess its energy-efficiency objective.
- Motivation: UCCF MMIMO can improve network Energy Efficiency (EE) by addressing underutilized but energy-consuming BSs in network-centric single-BS serving.The network-centric approach typically serves each UE from the BS providing the highest received power.
- Limitations of prior work: Existing analytical solutions for Serving Cluster Formulation (SCF) often rely on simplified models that omit realistic multi-carrier scheduling and radio-front-end characteristics.Examples include single-carrier operation, Rayleigh channels, omission of space-time-frequency Radio Resource Scheduler (RRS), and neglected nonlinear radio-front-end behavior.
- Proposed approach: The paper proposes an LLM-based AI agent to perform SCF in UCCF MMIMO and improve network EE through autonomous interaction with the environment.The proposal specifically targets serving-cluster decisions in the UCCF MMIMO network.
- Evaluation: The proposed solution is evaluated with a realistic system-level UCCF-MMIMO simulator based on a 3D Ray Tracer.The simulator considers OFDMA and a dedicated space-time-frequency RRS, among other system features.
II. Energy-Efficient, LLM-based SCF
The proposed LLM-based AI agent performs dynamic serving cluster formulation in a UCCF MMIMO network to improve energy efficiency. Hosted by the CPU, it operates through reinforcement-learning state recognition, action selection, reward observation, and state updates.
- The CPU coordinates BS transmissions and hosts the LLM-based AI agent, which performs energy-efficient serving cluster formulation.The network includes multiple MMIMO BSs attached to a Central Processing Unit.
- The agent operates in a reinforcement-learning cycle that recognizes the state, takes action using past experience, observes the reward, and recognizes a new state.
- The state is a UE-by-BS matrix of downlink received powers, enabling the agent to identify whether UEs should be served by one or multiple BSs.
- The action is a UE-length integer vector specifying how many highest-received-power BSs should serve each UE.
- The reward is network EE, defined as median user throughput divided by total power consumption to maintain greater fairness for users with worse radio conditions.
A. System Prompt
The system prompt defines an AI operator that selects how many coordinated base stations serve each user from received-power arrays and past EE experience. It requires complete, index-aligned list outputs and encourages exploration while adapting assignments based on observed EE.
- Network role and serving principle: The AI operator controls a 6G network with three base stations under the user-centric cell-free paradigm, where users may be served by multiple coordinated base stations.Serving multiple base stations is favored when received powers are similar; otherwise, serving with only one may be preferable.
- Input and output formulation: The prompt maps each user’s received-power row to an integer specifying the number of base stations that should serve that user.The input is a 2D received-power array in dBm with users as rows and base stations as columns.
- Experience-guided adaptation: Without past experience, assignments start at one base station per user; with experience, the agent increases assignments when this improves EE and explores new options.When recent EE is similar over the latest three attempts, the prompt permits minor random modifications; after more than 40 samples, it selects the experienced assignment list with the highest EE.
- Output constraints: The response must include every input user, have length equal to the number of users, and contain only the assignment list without explanations.Users are indexed from zero in the output specification.
B. User Prompt
The user prompt provides a representative serving and throughput history with network energy efficiency, then asks the model to choose serving-base-station counts from a current received-power array to maximize network EE. The response must follow strict output and per-user serving constraints.
- User Prompt: The prompt supplies a past example containing serving-base-station counts, per-user throughput values, and network EE of 1 kbit/J.The example lists serving counts [1, 1, 1, 2, 1] and throughputs [1, 1.5, 0.5, 2, 1].
- User Prompt: Given the current user–base-station received-power array, the model must decide how many base stations serve each user to maximize network EE.The array contains received powers in dBm for five users across three base stations.
- User Prompt: The output must contain exactly five entries, include every user, return only the specified list, and assign each user at most three serving base stations.These constraints define the required output size and per-user serving limit.
III. Simulation Studies
The simulation evaluates the proposed LLM-based AI agent in a realistic system-level UCCF MIMO simulator and compares GPT-5, GPT-4o, and GPT-4o-mini against the Network Centric approach. After stabilization, GPT-4o reaches about 90 kbit/J and provides up to 32% EE gain over the Network Centric baseline, while throughput improvements vary by GPT model.
- Simulation setup: The evaluation uses a system-level UCCF MIMO simulator with a realistic 3D Ray Tracer channel model and dedicated space-time-frequency resource scheduling.The simulator also incorporates OFDM-based MMIMO modeling, power consumption, and nonlinear transmitter-front-end characteristics with Rapp-modeled power-amplifier distortion.
- Compared methods: The proposed AI agent is evaluated with GPT-5, GPT-4o, and GPT-4o-mini, alongside comparisons against the Network Centric approach.The simulations used the latest versions of the GPT models available in November 2025.
- Energy efficiency: 90 kbit/J: GPT-4o achieves about this EE during the last 200 slots, compared with about 80 kbit/J for GPT-5 after stabilization.The AI agent gains knowledge and explores potential serving clusters over time before the solutions stabilize.
- Energy efficiency: 32%: GPT-4o achieves up to this EE gain compared with the Network Centric approach.This comparison is reported for the stabilized results in the last 200 slots.
- Throughput: 4% and 0.5%: GPT-5 and GPT-4o improve the 10th-percentile UE throughput by about these amounts, respectively, compared with the Network Centric approach.The throughput statistics are calculated from UE throughputs averaged over the last 200 time slots.
IV. Conclusion
The proposed LLM-based AI agent dynamically formulates serving clusters and improves UCCF MMIMO network energy efficiency by up to 32%. In realistic simulations, it outperformed state-of-the-art approaches in EE without reducing QoS for UEs in the worst or best radio conditions, with GPT-4o showing the highest potential.
- Conclusion: 32%: The proposed LLM-based AI agent improves UCCF MMIMO network EE through dynamic SCF.The reported improvement is an upper bound from the simulation study.
- Conclusion: The AI agent outperformed state-of-the-art approaches in EE in a realistic simulation environment.The evaluation used a realistic environment to assess the proposed approach.
- Conclusion: The EE improvement did not decrease QoS for UEs with the worst or best radio conditions.The conclusion explicitly covers both extremes of UE radio conditions.
- Conclusion: GPT-4o showed the highest potential among the evaluated LLM models.The study found that AI-agent performance depends on the LLM model used.